A Robust Deep Feature Extraction Method for Human Activity Recognition Using a Wavelet Based Spectral Visualisation...

Area of Science:

• Computer Science
• Biomedical Engineering
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for smart homes and healthcare, but privacy concerns limit vision-based methods.
• Wearable sensors offer a privacy-preserving alternative, yet extracting features from 1D data is challenging.

Purpose of the Study:

• To develop a novel method for extracting deep features from 1D wearable sensor data for improved HAR.
• To leverage time-frequency analysis and deep learning for accurate and efficient activity recognition.

Main Methods:

• Converted 1D multi-sensor data (accelerometer, gyroscope) into spectral images using continuous wavelet transform (scalograms).
• Employed deep learning models (CNN, MobileNetV3, ResNet, GoogleNet) to extract features from these spectral images.
• Classified activities using extracted features with a conventional classifier (softmax).

Main Results:

• Achieved high accuracy in HAR: 98.4% on the SisFall dataset and 98.1% on the PAMAP2 dataset.
• The proposed method, utilizing Morlet wavelet and ResNet-101, outperformed existing state-of-the-art algorithms.
• Demonstrated the effectiveness of time-frequency analysis for feature extraction in wearable sensor data.

Conclusions:

• The proposed time-frequency analysis combined with deep learning offers a robust and accurate solution for wearable sensor-based HAR.
• This approach enhances privacy in smart environments and has significant implications for Ambient Assisted Living (AAL).